IImageSafetyClassifier<T>

Interfaces ยท AiDotNet.Safety.Image

Interface for image safety classifiers that detect NSFW, violent, or otherwise harmful images.

Implements: IImageSafetyModule<T>

Known implementations: ImageSafetyClassifierBase<T>

For Beginners

An image safety classifier looks at an image and determines whether it contains harmful content. Different implementations use different approaches โ€” CLIP embeddings, Vision Transformers, scene graph analysis, or an ensemble of all three.

How It Works

Image safety classifiers analyze image tensors and assign per-category safety scores for categories including sexual content, violence, self-harm, hate symbols, drugs, child exploitation, shocking content, and dangerous activities.

References:

  • UnsafeBench: 11 categories, GPT-4V achieves top F1 (2024, arxiv:2405.03486)
  • USD: Scene-graph-based NSFW detection (USENIX Security 2025)
  • Sensitive image classification via Vision Transformers (2024, arxiv:2412.16446)

Methods

Method Summary
IReadOnlyDictionary<SafetyCategory, double> GetCategoryScores(Tensor<T> image) Gets per-category safety scores for the given image.